This analysis looks at the field data collected for the following 2017 MLRA project: - MLRA 111B - Glynwood B-slope Erosion; Northeastern IN
Spatially disaggregate the existing SSURGO polygons for Glynwood B-slope map units using ArcSIE, in order to separate different soil erosion phases. The current SSURGO maps join issues at the SSA boundaries, due to different erosion phases. This project is deemed relevant due to the current interest in Soil Health. Distinguishing the difference in erosion phases may have minimal impact on the majority of soil interpretations, but is believed to be significant in distinguishing crop yields.
library(aqp)
library(soilDB)
library(reshape2)
library(ggplot2)
library(gridExtra)
library(knitr)
library(cluster)
library(caret)
library(party)
library(vegan)
library(rgdal)
library(sp)
library(mapview)
gp <- fetchNASIS()
s <- site(gp)[c("peiid", "pedon_id", "taxonname", "x", "y", "describer", "erocl")]
h <- horizons(gp)
vars <- c("peiid")
A_vars <- c("hzname", "hzdept", "clay", "texture", "m_hue", "m_value", "m_chroma")
Bt_vars <- c("hzdept", "clay", "texture", "m_value", "m_chroma")
carb_vars <- c("hzdept", "effervescence")
solum_vars <- c("hzdept")
h2 <- by(h, h[vars], function(x) data.frame(
x[vars][1, ],
x[A_vars][1, ],
x[grepl("Bt", x$hzname), Bt_vars, ][1, ],
x[x$effervescence %in% c("strong", "violent"), carb_vars][1,],
x[grepl("^C|^2C|^3C", x$hzname), solum_vars][1]
))
h2 <- do.call("rbind", h2)
names(h2) <- c(vars, "hzname", "A_hzthk", "claytotest", "texture", "mxhue", "mxvalue", "mxchroma", "hzthk", "firstbtclay", "firstbttexture", "firstbtmxvalue", "firstbtmxchroma", "CaCO3Dp", "effervescence", "SolumDp")
names(s)[c(2:5, 7)] <- c("upedonid", "soilname", "long", "lat", "EroClassFD")
gw <- merge(s, h2, by = "peiid", all.x = TRUE)
gw <- gw[order(gw$upedonid), ]
gw <- within(gw,{
CaCO3Dp[is.na(CaCO3Dp)] <- 200
SolumDp[is.na(SolumDp)] <- 200
})
# projectmapunit data from NASIS
project <- get_projectmapunit_from_NASIS()
project_nodups <- project[!duplicated(project$nationalmusym), c("nationalmusym", "muname")]
# MUPOLYGONs for the Project
gw_pol <- readOGR(dsn = paste0(ownCloud, "glynwood.shp"), layer = "glynwood")
# Soil Survey Areas
ssa <- readOGR(dsn = paste0(geodata, "soils/soilsa_a_nrcs.shp"), layer = "soilsa_a_nrcs")
# Series Extent of Glynwood from SoilWeb
gw_series <- seriesExtent("Glynwood")
# field Data
gw2 <- read.csv(paste0(ownCloud, "Pts_gnbero_27Jan17.csv"))
vars <- c("upedonid", "EroClassSIE", "relpos", "SlopeSIE", "wetness", "PlanCrv", "ProfCrv", "maxcrv", "mincrv")
gw <- merge(gw, gw2[vars], by = "upedonid", all.x = TRUE)
gw <- gw[complete.cases(gw[c("lat", "long")]), ]
gw_sp <- gw
coordinates(gw_sp) <- ~ long + lat
proj4string(gw_sp) <- CRS("+init=epsg:4326")
gw_sp <- spTransform(gw_sp, CRS(proj4string(gw_pol)))
# spatial overlay field data with mupolygons and merge with nasis mapunits
vars <- c("AREASYMBOL", "nationalmu")
gw <- cbind(gw, over(gw_sp, gw_pol)[vars])
gw <- merge(gw, project_nodups, by.x = "nationalmu", by.y = "nationalmusym", all.x = TRUE, sort = FALSE)
# Extract erosion phases from NASIS and combine component and phase
ero_labels <- c("undisturbed", "slight", "moderate", "severe")
gw <- within(gw, {
EroClassFD = factor(EroClassFD, levels = 0:3, labels = ero_labels)
EroClassSIE = factor(EroClassSIE, levels = 0:3, labels = ero_labels)
EroClassFD2 = ifelse(EroClassFD == "severe", "severe", "slight")
EroClassNASIS = NA
EroClassNASIS[grepl("eroded", muname)] = "eroded"
EroClassNASIS[grepl("sev.|severely", muname)] = "sev.eroded"
EroClassNASIS[!grepl("eroded", muname)] = "non.eroded"
soilname2 = soilname
soilname2 = ifelse(soilname2 %in% c("Glynwood", "Morley", "Shinrock", "Rawson", "Mississinewa"), "Glynwood", soilname2)
soilname2 = ifelse(soilname2 %in% c("Blount", "Elliott"), "Blount", soilname2)
soilname2 = ifelse(soilname2 %in% c("Pewamo", "Pandora", "Mermill"), "Pewamo", soilname2)
soilname3 = paste0(soilname2, ifelse(soilname2 == "Glynwood", paste0("-", EroClassFD2), ""))
})
gw <- transform(gw,
rgb = munsell2rgb(mxhue, mxvalue, mxchroma, return_triplets = TRUE)
)
gw_sp <- gw
coordinates(gw_sp) <- ~ long + lat
proj4string(gw_sp) <- CRS("+init=epsg:4326")
gw_sp <- spTransform(gw_sp, CRS(proj4string(gw_pol)))
The geodata from the Glynwood points was extracted from several rasters at various resolutions. The data using to generate the ArcSIE model came from a DEM with a resolution of 15-feet. The other used came from the 10-meter USGS NED, which was primarily resampled from LiDAR.
# Extract data from rasters
library(raster)
# NW files
fd <- paste0(geodata, "project_data/11FIN/PointDataEval/")
dd <- c("slope10",
"procrv10",
"plncrv10",
"maxcrv10",
"mincrv10",
"relpos_r5",
"wetness_mp"
)
fp <- paste0(fd, "Mosaic_NW_pts/Derivatives/", dd, "/", "w001001.adf")
rs <- stack(fp)
names(rs) <- dd
proj4string(rs) <- CRS("+init=epsg:2965")
gd_nw <- extract(rs, gw_sp, df = TRUE, sp = TRUE)@data
gd_nw <- subset(gd_nw, !is.na(slope10))
# SE files
fp <- paste0(fd, "Mosaic_SE_pts/Derivatives/", dd, "/", "w001001.adf")
rs <- stack(fp)
names(rs) <- dd
proj4string(rs) <- CRS("+init=epsg:2965")
gd_se <- extract(rs, gw_sp, df = TRUE, sp = TRUE)@data
gd_se <- subset(gd_se, !is.na(slope10))
gd15ft <- rbind(gd_nw, gd_se)
rm(gd_nw, gd_se)
write.csv(gd15ft, file = paste0(ownCloud, "geodata_15ft_extract.csv"))
# Region 11 files
subset_rasters <- function(input, output) {
cat(paste0(input, "\n"))
gdal_translate(
src_dataset = input,
dst_dataset = output,
projwin = c(bb[1,1], bb[2,2], bb[1,2], bb[2,1]),
of = "GTiff",
a_nodata = -99999,
overwrite = TRUE,
verbose = TRUE
)
}
warp_rasters <- function(input, output){
cat(paste0(input,"\n"))
gdalwarp(
srcfile = input,
dstfile = output,
te = bb,
s_srs = CRSargs(CRS("+init=epsg:5070")),
t_srs = CRSargs(CRS("+init=epsg:5070")),
r = "bilinear",
tr = c(10, 10),
of = "GTiff",
overwrite = TRUE,
verbose = TRUE
)
}
fd <- paste0(geodata, "project_data/11FIN/sdat/")
dd2 <- c("ned10m_11FIN.sdat",
"ned10m_11FIN_aspect5.sdat",
"ned10m_11FIN_slope5.sdat",
"ned10m_11FIN_cupro5.sdat",
"ned10m_11FIN_cutan5.sdat",
"ned30m_11FIN_mvalleys.sdat",
"ned30m_11FIN_wetness.sdat",
"ned30m_11FIN_z2stream.sdat",
"nlcd30m_11FIN_lulc2011.sdat"
)
dd_names <- c("elev", "aspect5", "slope5", "kp5", "kt5", "mvalley", "wetness2", "z2streams", "lulc")
dd <- paste0(fd, dd2)
input <- dd
output <- paste0(geodata, "project_data/11FIN/glynwood/", gsub(".sdat", ".tif", dd2))
bb <- bbox(gw_sp <- spTransform(gw_sp, CRS("+init=epsg:5070")))
mapply(subset_rasters, input, output)
mapply(subset_rasters, input = paste0(geodata, "project_data/11FIN/nlcd30m_11FIN_lulc2011.tif"), output = paste0(geodata, "project_data/11FIN/glynwood/nlcd30m_11FIN_lulc2011.tif"))
mapply(warp_rasters, input = output, output = gsub(".tif", "2.tif", gsub("30m", "10m", output)))
dd <- output
rs10m <- stack(dd[grepl("10m", dd)])
names(rs10m) <- dd_names[1:5]
rs30m <- stack(dd[grepl("30m", dd)])
names(rs30m) <- dd_names[6:9]
rs10m <- stack(gsub(".tif", "2.tif", gsub("30m", "10m", output)))
names(rs10m) <- dd_names
gd10m <- as.data.frame(extract(rs10m, gw_sp, df = TRUE, sp = TRUE))
gd30m <- as.data.frame(extract(rs30m, gw_sp, df = TRUE))
gw <- cbind(gd10m, gd30m[, -1])
rm(gd10m, gd30m)
# Save data
save(gw, gw_sp, gw_pol, gw_series, ssa, ero_labels, file = paste0(ownCloud, "Pts_gnbero_27Jan17_geodata2.RData"))
# Load cached dataset
load(paste0(ownCloud, "Pts_gnbero_27Jan17_geodata2.RData"))
ssa <- subset(ssa, areasymbol %in% unique(gw$AREASYMBOL))
ssa <- spTransform(ssa, CRS("+init=epsg:4326"))
# Create interactive map
mapView(gw_series) + ssa + gw_sp
vals2 <- c("EroClassFD", "EroClassNASIS", "nationalmu", "AREASYMBOL")
gw_sub <- gw[vals2]
# Frequency of field observation vs map unit
# Duplicate the data for each REASYMBOL and relabel MLRA
gw_sub2 <- by(gw_sub, gw_sub$nationalmu, function(x) {
x[vals2][1, ]
x[, "AREASYMBOL"] <- "MLRA"
return(x)
})
gw_sub2 <- do.call("rbind", gw_sub2)
gw_sub <- rbind(gw_sub, gw_sub2)
gw_sub$natmuSsaEro <- with(gw_sub,
paste0(nationalmu, "-", AREASYMBOL, "-", EroClassNASIS)
)
test <- xtabs(~ natmuSsaEro + EroClassFD, data = gw_sub)
kable(test, caption = "Frequence by MUSYM-SSA-EROSION")
| undisturbed | slight | moderate | severe | |
|---|---|---|---|---|
| 2t6ll-IN009-sev.eroded | 13 | 14 | 17 | 3 |
| 2t6ll-IN053-sev.eroded | 4 | 11 | 17 | 11 |
| 2t6ll-IN075-sev.eroded | 6 | 8 | 4 | 5 |
| 2t6ll-IN179-sev.eroded | 1 | 9 | 6 | 10 |
| 2t6ll-MLRA-sev.eroded | 24 | 42 | 44 | 29 |
| 2t6lm-IN009-sev.eroded | 2 | 11 | 10 | 1 |
| 2t6lm-IN053-sev.eroded | 2 | 4 | 8 | 11 |
| 2t6lm-IN075-sev.eroded | 3 | 1 | 11 | 9 |
| 2t6lm-IN179-sev.eroded | 9 | 8 | 0 | 13 |
| 2t6lm-MLRA-sev.eroded | 16 | 24 | 29 | 34 |
| 2v4bn-IN069-eroded | 4 | 3 | 6 | 4 |
| 2v4bn-IN179-eroded | 0 | 3 | 4 | 6 |
| 2v4bn-MLRA-eroded | 4 | 6 | 10 | 10 |
| 2v4bp-IN179-eroded | 0 | 3 | 0 | 2 |
| 2v4bp-MLRA-eroded | 0 | 3 | 0 | 2 |
| 5jjt-IN035-sev.eroded | 1 | 1 | 2 | 9 |
| 5jjt-MLRA-sev.eroded | 1 | 1 | 2 | 9 |
| NA-NA-non.eroded | 6 | 9 | 10 | 34 |
kable(round(prop.table(test, 1) * 100), caption = "Percent by MUSYM-SSA-EROSION")
| undisturbed | slight | moderate | severe | |
|---|---|---|---|---|
| 2t6ll-IN009-sev.eroded | 28 | 30 | 36 | 6 |
| 2t6ll-IN053-sev.eroded | 9 | 26 | 40 | 26 |
| 2t6ll-IN075-sev.eroded | 26 | 35 | 17 | 22 |
| 2t6ll-IN179-sev.eroded | 4 | 35 | 23 | 38 |
| 2t6ll-MLRA-sev.eroded | 17 | 30 | 32 | 21 |
| 2t6lm-IN009-sev.eroded | 8 | 46 | 42 | 4 |
| 2t6lm-IN053-sev.eroded | 8 | 16 | 32 | 44 |
| 2t6lm-IN075-sev.eroded | 12 | 4 | 46 | 38 |
| 2t6lm-IN179-sev.eroded | 30 | 27 | 0 | 43 |
| 2t6lm-MLRA-sev.eroded | 16 | 23 | 28 | 33 |
| 2v4bn-IN069-eroded | 24 | 18 | 35 | 24 |
| 2v4bn-IN179-eroded | 0 | 23 | 31 | 46 |
| 2v4bn-MLRA-eroded | 13 | 20 | 33 | 33 |
| 2v4bp-IN179-eroded | 0 | 60 | 0 | 40 |
| 2v4bp-MLRA-eroded | 0 | 60 | 0 | 40 |
| 5jjt-IN035-sev.eroded | 8 | 8 | 15 | 69 |
| 5jjt-MLRA-sev.eroded | 8 | 8 | 15 | 69 |
| NA-NA-non.eroded | 10 | 15 | 17 | 58 |
Several of counties phased severely eroded, are not dominanted by field observations classified as severely eroded.
cm <- confusionMatrix(data = gw$EroClassSIE, reference = gw$EroClassFD)
print(cm)
## Confusion Matrix and Statistics
##
## Reference
## Prediction undisturbed slight moderate severe
## undisturbed 0 0 0 0
## slight 9 6 10 9
## moderate 27 34 51 37
## severe 12 31 21 45
##
## Overall Statistics
##
## Accuracy : 0.3493
## 95% CI : (0.2947, 0.407)
## No Information Rate : 0.3116
## P-Value [Acc > NIR] : 0.09322
##
## Kappa : 0.0862
## Mcnemar's Test P-Value : 1.115e-14
##
## Statistics by Class:
##
## Class: undisturbed Class: slight Class: moderate
## Sensitivity 0.0000 0.08451 0.6220
## Specificity 1.0000 0.87330 0.5333
## Pos Pred Value NaN 0.17647 0.3423
## Neg Pred Value 0.8356 0.74806 0.7832
## Prevalence 0.1644 0.24315 0.2808
## Detection Rate 0.0000 0.02055 0.1747
## Detection Prevalence 0.0000 0.11644 0.5103
## Balanced Accuracy 0.5000 0.47891 0.5776
## Class: severe
## Sensitivity 0.4945
## Specificity 0.6816
## Pos Pred Value 0.4128
## Neg Pred Value 0.7486
## Prevalence 0.3116
## Detection Rate 0.1541
## Detection Prevalence 0.3733
## Balanced Accuracy 0.5880
test <- as.data.frame(cm$table)
ggplot(test, aes(x = Reference, y = Freq, fill = Prediction)) +
geom_bar(stat = "identity") +
coord_flip()
The accuracy of the current ArcSIE model appears to be low, according to several metrics. The positive predictive value for the severe class is < 50%.
soil_vals <- c("hzthk", "SolumDp", "CaCO3Dp", "claytotest", "firstbtclay", "mxvalue", "mxchroma")
geo_vals1 <- c("SlopeSIE", "ProfCrv", "PlanCrv", "relpos", "wetness")
geo_vals2 <- c("slope5", "kt5", "kp5", "z2streams", "wetness2", "mvalley")
vals <- c(soil_vals, geo_vals1, geo_vals2)
gw <- gw[complete.cases(gw[c("EroClassFD", soil_vals)]), ]
gw_lo1 <- melt(gw, id.vars = "EroClassFD", measure.vars = vals)
gw_lo2 <- melt(gw, id.vars = "EroClassSIE", measure.vars = vals)
names(gw_lo1)[1] <- "EroClass"
gw_lo1$method <- "FD"
names(gw_lo2)[1] <- "EroClass"
gw_lo2$method <- "SIE"
gw_lo <- rbind(gw_lo1, gw_lo2)
gw_lo <- subset(gw_lo, !is.na(EroClass))
gw_lo <- na.exclude(gw_lo)
ggplot(gw_lo, aes(x = EroClass, y = value)) +
geom_boxplot() +
facet_wrap(~ paste(variable, method), scales="free", ncol = 4) +
coord_flip()
An exploratory analysis shows a considerable amount of overlap exists between the field determined (FD) erosion classes and measurable soil properties. In comparison the FD and SIE (Soil Inference Engine) erosion classes show different patterns within the boxplots, further suggesting that the SIE classes aren’t capturing the field observations accurately. The most important feature to highlight is that the trends between the SIE classes and digital elevation model (DEM) derivatives (i.e. slope) don’t match those observed for the FD classes. This mismatch suggests that the membership functions for the SIE classes are a poor fit, and should be redefined to more accurately represent the relationship between the FD classes and DEM derivatives.
soil_vals2 <- c("hzthk", "SolumDp", "CaCO3Dp", "claytotest", "firstbtclay") # excluded color, only observed a narrow range thus small differences swamp everthing else
vals <- c(soil_vals2)
test <- gw[, vals]
test_d <- daisy(scale(test), metric = "gower")
test_mds <- metaMDS(test_d, distance = "gower", autotransform = FALSE, trace = FALSE)
test_pts <- cbind(as.data.frame(test_mds$points), EroClassFD = gw$EroClassFD)
g1 <- ggplot(gw, aes(x = hzthk, y = SolumDp, color = EroClassFD)) +
geom_point(cex = 2, alpha = 0.75) +
theme(aspect.ratio = 1)
g2 <- ggplot(test_pts, aes(x = MDS1, y = MDS2, color = EroClassFD)) +
geom_point(cex = 2, alpha = 0.75) +
theme(aspect.ratio = 1)
grid.arrange(g1, g2, ncol = 2)
According to the scatterplot above it appears that only the severe and slight classes are separatable. The moderate erosion class seems to overlap the most with slight. The overlap in the FD classes is likely due to bias within and between the soil scientists who collected the data. Both the 15-feet and 10-meter DEM derivatives were evaluated, but the results are similar.
test <- subset(gw, !is.na(EroClassFD))
test_ct <- ctree(EroClassFD ~ ., data = test[, c("EroClassFD", soil_vals)])
plot(test_ct)
cm <- confusionMatrix(data = predict(test_ct, type = "response"), reference = test$EroClassFD)
print(cm)
## Confusion Matrix and Statistics
##
## Reference
## Prediction undisturbed slight moderate severe
## undisturbed 37 5 1 0
## slight 8 62 3 1
## moderate 5 14 66 4
## severe 0 2 23 104
##
## Overall Statistics
##
## Accuracy : 0.803
## 95% CI : (0.7563, 0.8442)
## No Information Rate : 0.3254
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 0.729
## Mcnemar's Test P-Value : NA
##
## Statistics by Class:
##
## Class: undisturbed Class: slight Class: moderate
## Sensitivity 0.7400 0.7470 0.7097
## Specificity 0.9789 0.9524 0.9050
## Pos Pred Value 0.8605 0.8378 0.7416
## Neg Pred Value 0.9555 0.9195 0.8902
## Prevalence 0.1493 0.2478 0.2776
## Detection Rate 0.1104 0.1851 0.1970
## Detection Prevalence 0.1284 0.2209 0.2657
## Balanced Accuracy 0.8595 0.8497 0.8073
## Class: severe
## Sensitivity 0.9541
## Specificity 0.8894
## Pos Pred Value 0.8062
## Neg Pred Value 0.9757
## Prevalence 0.3254
## Detection Rate 0.3104
## Detection Prevalence 0.3851
## Balanced Accuracy 0.9218
An analysis of the EroClassFD above with a classification tree is an attempt to discern the hierachical structuce within the data. The results show Ap thickness (hzthk) and clay content (claytotest) are the first splits. The trees structure follows the logic described in the erosion indicators guide developed for this project. The overall accuracy for the tree is 0.8.
In order to see if more separation can be achieved amongst the erosion classes a hierachical classifition was peformed. Four unsupervised classes were chosen and manually matched to the FD classes.
test_c <- hclust(test_d, method = "ward")
plot(test_c, labels = gw$upedonid)
rect.hclust(test_c, k = 4)
clusters <- cbind(gw,
test_pts[, 1:2],
clusters = factor(cutree(test_c, k = 4),
levels = c(2, 3, 1, 4),
labels = ero_labels
)
)
clusters <- cbind(gw,
test_pts[, 1:2],
clusters = factor(cutree(test_c, k = 4),
levels = c(4, 1, 2, 3),
labels = ero_labels
)
)
xtabs(~ EroClassFD + clusters, data = clusters)
## clusters
## EroClassFD undisturbed slight moderate severe
## undisturbed 31 14 5 0
## slight 11 54 18 0
## moderate 5 28 43 17
## severe 0 5 34 70
g1 <- ggplot(clusters, aes(x = MDS1, y = MDS2, col = EroClassFD)) +
geom_point(cex = 2, alpha = 0.75) +
theme(aspect.ratio = 1)
g2 <- ggplot(clusters, aes(x = MDS1, y = MDS2, col = clusters), main = "test") +
geom_point(cex = 2, alpha = 0.75) +
theme(aspect.ratio = 1)
grid.arrange(g1, g2, ncol = 2)
In comparison the hierarchical clusters have less overlap when viewed along the multidimensional scaled axes, but still does not seem to separate the moderate class.
gw_lo3 <- melt(clusters, id.vars = "clusters", measure.vars = c(soil_vals, geo_vals2))
names(gw_lo3)[1] <- "EroClass"
gw_lo3$method <- "clusters"
gw_lo1 <- subset(gw_lo1, ! variable %in% c("relpos", "wetness", "SlopeSIE"))
gw_lo <- rbind(gw_lo1, gw_lo3)
ggplot(gw_lo, aes(x = EroClass, y = value)) +
geom_boxplot() +
facet_wrap(~ paste(variable, method), scales="free", ncol = 4) +
coord_flip()
A comparison of the FD and cluster classes shows that the clusters do a good job replicating the patterns found in the boxplots.
test2 <- ctree(clusters ~ ., data = clusters[, c("clusters", soil_vals)])
plot(test2)
confusionMatrix(data = predict(test2, type = "response"), reference = clusters$clusters)
## Confusion Matrix and Statistics
##
## Reference
## Prediction undisturbed slight moderate severe
## undisturbed 47 0 0 0
## slight 0 91 3 1
## moderate 0 7 89 1
## severe 0 3 8 85
##
## Overall Statistics
##
## Accuracy : 0.9313
## 95% CI : (0.8988, 0.956)
## No Information Rate : 0.3015
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 0.9065
## Mcnemar's Test P-Value : NA
##
## Statistics by Class:
##
## Class: undisturbed Class: slight Class: moderate
## Sensitivity 1.0000 0.9010 0.8900
## Specificity 1.0000 0.9829 0.9660
## Pos Pred Value 1.0000 0.9579 0.9175
## Neg Pred Value 1.0000 0.9583 0.9538
## Prevalence 0.1403 0.3015 0.2985
## Detection Rate 0.1403 0.2716 0.2657
## Detection Prevalence 0.1403 0.2836 0.2896
## Balanced Accuracy 1.0000 0.9419 0.9280
## Class: severe
## Sensitivity 0.9770
## Specificity 0.9556
## Pos Pred Value 0.8854
## Neg Pred Value 0.9916
## Prevalence 0.2597
## Detection Rate 0.2537
## Detection Prevalence 0.2866
## Balanced Accuracy 0.9663
In comparision, the classification tree for the clusters splits primarily on the CaCO3 and solum depths, presumable due to the narrow range in Ap thickness.
Below several statistical models were evaluated to see if a more accurate model could be developed.
test3 <- ctree(EroClassFD ~ ., data = gw[, c("EroClassFD", geo_vals2)])
plot(test3)
cm_ct <- confusionMatrix(data = predict(test3, type = "response"), reference = gw$EroClassFD)
round(cm_ct$overall, 2)
## Accuracy Kappa AccuracyLower AccuracyUpper AccuracyNull
## 0.38 0.11 0.32 0.43 0.33
## AccuracyPValue McnemarPValue
## 0.03 0.00
test3 <- cforest(as.factor(EroClassFD) ~ ., data = gw[, c("EroClassFD", geo_vals2)])
varimp(test3)
## slope5 kt5 kp5 z2streams wetness2 mvalley
## 0.037024390 0.019056911 0.005479675 0.001138211 0.016569106 0.000699187
cm_cf <-confusionMatrix(data = predict(test3, type = "response", OOB = TRUE), reference = gw$EroClassFD)
round(cm_cf$overall, 2)
## Accuracy Kappa AccuracyLower AccuracyUpper AccuracyNull
## 0.37 0.14 0.32 0.43 0.33
## AccuracyPValue McnemarPValue
## 0.04 0.03
Neither a classification tree or forest were capiable of achieving a higher accuracy than the SIE model.
test4 <- ctree(clusters ~ ., data = clusters[, c("clusters", geo_vals2)])
plot(test4)
cm_ct <- confusionMatrix(data = predict(test4, type = "response"), reference = clusters$clusters)
round(cm_ct$overall, 2)
## Accuracy Kappa AccuracyLower AccuracyUpper AccuracyNull
## 0.38 0.12 0.32 0.43 0.30
## AccuracyPValue McnemarPValue
## 0.00 0.00
test4 <- cforest(clusters ~ ., data = clusters[, c("clusters", geo_vals2)])
varimp(test4)
## slope5 kt5 kp5 z2streams wetness2
## 0.0186016260 0.0343252033 0.0004227642 0.0017073171 0.0084390244
## mvalley
## 0.0131544715
cm_cf <- confusionMatrix(data = predict(test4, type = "response", OOB=TRUE), reference = clusters$clusters)
round(cm_cf$overall, 2)
## Accuracy Kappa AccuracyLower AccuracyUpper AccuracyNull
## 0.36 0.12 0.31 0.42 0.30
## AccuracyPValue McnemarPValue
## 0.01 0.39
Neither a classification tree or forest were capiable of achieving a higher accuracy than the SIE model.
Thus far efforts to model the erosion classes has been lackluster. This appears to be largely due to the overlap in the erosion classes and subtle relief. Given these challenges it is probably more realistic to focus on distinguishing the severely eroded class separately, and develop individual models for the minor components. ry and model the soil components and phases separately.
# create a logical variable for the soilname3 == "Glynwood-severe"
gw$gw_severe <- ifelse(gw$soilname3 == "Glynwood-severe", TRUE, FALSE)
# Random Forest
test4 <- cforest(as.factor(gw_severe) ~ elev + slope5 + kt5 + kp5 + wetness2 + mvalley + z2streams, data = gw)
sort(varimp(test4), decreasing = TRUE)
## slope5 kt5 wetness2 elev z2streams kp5
## 0.031560976 0.022000000 0.019560976 0.018926829 0.003707317 0.002552846
## mvalley
## 0.002520325
confusionMatrix(data = predict(test4, type = "response", OOB = TRUE), reference = gw$gw_severe, positive = "TRUE")
## Confusion Matrix and Statistics
##
## Reference
## Prediction FALSE TRUE
## FALSE 197 64
## TRUE 29 45
##
## Accuracy : 0.7224
## 95% CI : (0.6711, 0.7697)
## No Information Rate : 0.6746
## P-Value [Acc > NIR] : 0.0340262
##
## Kappa : 0.3103
## Mcnemar's Test P-Value : 0.0004225
##
## Sensitivity : 0.4128
## Specificity : 0.8717
## Pos Pred Value : 0.6081
## Neg Pred Value : 0.7548
## Prevalence : 0.3254
## Detection Rate : 0.1343
## Detection Prevalence : 0.2209
## Balanced Accuracy : 0.6423
##
## 'Positive' Class : TRUE
##
# Logisitic Regression
test3 <- glm(as.factor(gw_severe) ~ elev + slope5 + kt5, data = gw, family = "binomial", na.action = na.exclude)
confusionMatrix(data = predict(test3, type = "response") > 0.4, reference = gw$gw_severe, positive = "TRUE")
## Confusion Matrix and Statistics
##
## Reference
## Prediction FALSE TRUE
## FALSE 180 48
## TRUE 43 60
##
## Accuracy : 0.7251
## 95% CI : (0.6736, 0.7725)
## No Information Rate : 0.6737
## P-Value [Acc > NIR] : 0.02528
##
## Kappa : 0.3671
## Mcnemar's Test P-Value : 0.67499
##
## Sensitivity : 0.5556
## Specificity : 0.8072
## Pos Pred Value : 0.5825
## Neg Pred Value : 0.7895
## Prevalence : 0.3263
## Detection Rate : 0.1813
## Detection Prevalence : 0.3112
## Balanced Accuracy : 0.6814
##
## 'Positive' Class : TRUE
##
summary(test3)
##
## Call:
## glm(formula = as.factor(gw_severe) ~ elev + slope5 + kt5, family = "binomial",
## data = gw, na.action = na.exclude)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -1.5671 -0.8653 -0.6426 1.1272 2.1766
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -8.08574 1.93935 -4.169 3.06e-05 ***
## elev 0.02205 0.00683 3.229 0.00124 **
## slope5 0.39779 0.10284 3.868 0.00011 ***
## kt5 0.08349 0.02730 3.059 0.00222 **
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 418.06 on 330 degrees of freedom
## Residual deviance: 380.17 on 327 degrees of freedom
## (4 observations deleted due to missingness)
## AIC: 388.17
##
## Number of Fisher Scoring iterations: 3
gw$predicted <- predict(test3, type = "response") > 0.4
gw_lo1 <- melt(gw, id.vars = "gw_severe", measure.vars = vals)
gw_lo2 <- melt(gw, id.vars = "predicted", measure.vars = vals)
gw_lo2 <- na.exclude(gw_lo2)
names(gw_lo1)[1] <- "EroClass"
gw_lo1$method <- "FD"
names(gw_lo2)[1] <- "EroClass"
gw_lo2$method <- "GLM"
gw_lo <- rbind(gw_lo1, gw_lo2)
ggplot(gw_lo, aes(x = EroClass, y = value)) +
geom_boxplot() +
facet_wrap(~ paste(variable, method), scales="free", ncol = 4) +
coord_flip()
predfun <- function(model, data) {
v <- predict(model, data, type = "response")
cbind(
p = as.vector(v$fit)
)
}
r <- predict(rs10m, test3, fun = predfun, index = 1:2, progress = "text")
writeRaster(r[[1]], "C:/workspace/severe_erosion.tif", overwrite = TRUE, progress = "text")
r <-predict(rs10m, test4, type='response', progress='text')
writeRaster(r[[1]], "C:/workspace/severe_erosion_cf.tif", overwrite = TRUE, progress = "text")